DOI 10.1007/s00217-015-2537-4
ORIGINAL PAPER
Monovarietal extra‑virgin olive oil classification: a fusion
of human sensory attributes and an electronic tongue
Luís G. Dias1,2 · Nuno Rodrigues3 · Ana C. A. Veloso4,5 · José A. Pereira3 ·
António M. Peres6
Received: 27 May 2015 / Revised: 12 August 2015 / Accepted: 22 August 2015 / Published online: 3 September 2015 © Springer-Verlag Berlin Heidelberg 2015
sensory attributes) enabled 100 % leave-one-out cross-val-idation correct classification, improving the discrimination capability of the individual use of sensor profiles or sensory attributes (70 and 57 % leave-one-out correct classifica-tions, respectively). So, human sensory evaluation and elec-tronic tongue analysis may be used as complementary tools allowing successful monovarietal olive oil discrimination.
Keywords Single-cultivar extra-virgin olive oil · Sensory
analysis · Potentiometric electronic tongue · Linear multivariate methods · Simulated annealing algorithm Introduction
There is an increasing need for developing appropriate methodologies for guaranteeing olive oil commercial cate-gory (extra virgin, EVOO; virgin, VOO; and lampante olive oils, LOO), as well as for identifying geographical or olive
cultivar origin [1–3]. Olive oil organoleptic attributes [1,
4–6] and physicochemical parameters [7, 8] are important
for olive oil quality assessment and discrimination, avoid-ing the indiscriminate use of uncorrected definitions in
olive oil labels [1]. These analysis are time-consuming and
costly, requiring trained sensory panelists and high-skilled technicians, respectively. Moreover, sensory evaluation is mainly used to classify olive oils as LOO, VOO, or EVOO according to their positive and/or negative organoleptic sensations. In the last years, several analytical approaches,
ranging from expensive chromatographic [2, 6, 9–13]-,
DNA [14]-, or nondestructive spectroscopy [15, 16]-based
methods to fast and low-cost electrochemical devices [3,
13, 17–19], have been reported. These methods showed
sat-isfactory predictive performances regarding olive oil qual-ity assessment and classification, including the successful
Abstract Olive oil quality grading is traditionally
assessed by human sensory evaluation of positive and nega-tive attributes (olfactory, gustatory, and final olfactory– gustatory sensations). However, it is not guaranteed that trained panelist can correctly classify monovarietal extra-virgin olive oils according to olive cultivar. In this work, the potential application of human (sensory panelists) and artificial (electronic tongue) sensory evaluation of olive oils was studied aiming to discriminate eight single-cultivar extra-virgin olive oils. Linear discriminant, partial least square discriminant, and sparse partial least square discri-minant analyses were evaluated. The best predictive clas-sification was obtained using linear discriminant analysis with simulated annealing selection algorithm. A low-level data fusion approach (18 electronic tongue signals and nine * António M. Peres
peres@ipb.pt
1 Escola Superior Agrária, Instituto Politécnico de Bragança, Campus Santa Apolónia, Apartado 1172, 5301-855 Bragança, Portugal
2 CQ-VR, Centro de Química – Vila Real, University of Trás-os-Montes e Alto Douro, Apartado 1013, 5001-801 Vila Real, Portugal
3 CIMO - Mountain Research Centre, Escola Superior Agrária, Instituto Politécnico de Bragança, Campus Santa Apolónia, Apartado 1172, 5301-855 Bragança, Portugal
4 Instituto Politécnico de Coimbra, ISEC, DEQB, Rua Pedro Nunes, Quinta da Nora, 3030-199 Coimbra, Portugal 5 CEB - Centre of Biological Engineering, University
of Minho, Campus de Gualtar, 4710-057 Braga, Portugal 6 LSRE - Laboratory of Separation and Reaction Engineering -
Associate Laboratory LSRE/LCM, Escola Superior Agrária, Instituto Politécnico de Bragança, Campus Santa Apolónia, Apartado 1172, 5301-855 Bragança, Portugal
discrimination of EVOO, VOO, and LOO; the identifica-tion of monovarietal EVOO and their geographical origin as well as the identification of potential putative markers
for the recognition of olive oil adulteration [2]. Less
atten-tion has been given to the possibility of using sensory data for classifying monovarietal EVOO according to its olive
cultivar. In a previous work, Dias et al. [3] showed that an
electronic tongue (E-tongue) could be employed to dis-criminate three Portuguese or three Spanish monovarietal EVOOs, from two crop years, according to olive cultivar (100 % and 98 % of correct classifications, respectively). However, when the six single-cultivar EVOOs were evalu-ated simultaneously, the correct classification percentage dramatically decreased to 43 %. This drawback only over-come when EVOO samples were split by production year, which enabled to reduce data variability within each year group, allowing to increase the correct classification rate of
nine single-cultivar EVOOs to 93–95 % [20].
On the other hand, some research works aiming for the improvement of olive oil quality assessment and clas-sification according to olive cultivar or geographical ori-gin reported the successful use of data fusion approaches, which merged data from different instrumental analytical methods, such as headspace mass spectrometry (electronic nose), ultraviolet–visible and near-infrared spectroscopy
[21]; near- and mid-infrared spectroscopy [1, 22];
ultravio-let–visible, near- and mid-infrared spectroscopy together with fatty acid composition profile from gas
chromatog-raphy analysis [23]; electronic nose and tongue [19]; and
visible spectroscopic fingerprints and chemical
descrip-tors [24]. Indeed, data fusion enabled to obtain compatible
measurements originated from different sources [25]. Data
fusion approaches, with different abstraction levels (low, mid, or high level), were effectively applied to differenti-ate or classify different food matrices or their attributes,
namely beers [25, 26], snack foods [27], fruit juices [28],
and black teas [29].
In this work and for the first time, the feasibility of a low-level data fusion approach, combining E-tongue poten-tiometric signals and olfactory and/or gustatory sensations,
was studied aiming to improve the discrimination of high-value monovarietal EVOO according to olive cultivar. Signal profiles were recorded using an E-tongue, with cross-sensi-tivity nonspecific lipid membranes as previously described
[3]. The single-cultivar EVOO organoleptic profiles were
evaluated by four trained sensory panelists. Eight Spanish monovarietal EVOOs were used (cvs. Arbosana, Arroniz, Cornicabra, Frantoio, Manzanilla, Redondilla, Royuela, and Zorzal). Four supervised linear multivariate models were tested to establish EVOO classification according to olive cultivar based on the independent variables, potentiometric sensor signals, and/or sensory descriptors. As previously
reported [3], hydroethanolic extracts of EVOO, potentially
rich in polar compounds that may be correlated with bitter-ness, astringency, and pungency attributes, were analyzed to overcome the difficulty of carrying out electrochemical
assays in viscous non-conductive liquids [30]. In summary,
this work aimed to demonstrate the complementary capabil-ities of human and artificial sensory analysis.
Materials and methods
Olive oil samples
Eight monovarietal Spanish EVOOs (cv. Arbosana (ARBO), cv. Arroniz (ARR), cv. Cornicabra (COR), cv. Frantoio (FRA), cv. Manzanilla (MAN), cv. Redondilla (RED), cv. Royuela (ROY), and cv. Zorzal (ZOR)), pro-duced at the north of Spain (Valladolid region), were stud-ied. These traditional olive varieties were selected due to their recognized high-quality, peculiar, and differentiated organoleptic characteristics. Moreover, the capability of discriminating these EVOOs has became quite relevant since some of these olive cultivars are usually produced in intensive olive groves (e.g., ARBO and FRA) and so can-not be used to obtain Protected Denomination of Origin olive oils. In total, 22 different single-cultivar EVOOs were obtained directly from olive oil certified producers during
2012 and 2013 (Table 1). For each one of the EVOOs, two
Table 1 Sample details of the Spanish monovarietal EVOO obtained in Valladolid region
Varietal (cvs.) Number of EVOOs Number of samples (bottles from different production lots)
Production year(s)
Arbosana (ARBO) 4 8 2012 and 2013
Arroniz (ARR) 2 4 2012
Cornicabra (COR) 4 8 2012 and 2013
Frantoio (FRA) 4 8 2012 and 2013
Manzanilla (MAN) 2 4 2012
Redondilla (RED) 2 4 2012
Royuela (ROY) 2 4 2013
bottles of different production lots were analyzed, totaling 44 independent olive oil samples. Olive oils were packed and stored in dark at −20 °C for a 24-h period after their production in olive mills with a two-phase extraction
pro-cess and kept in those conditions until further analysis [3].
Sensory analysis
Olive oil samples were subjected to organoleptic assess-ment following the methods and standards adopted by the International Olive Council (IOC or COI) for sensory analysis of olive oils, namely COI/T.20/Doc. No. 15/Rev.
6 [31] and COI/T.30/Doc. No. 17 [32]. Each sample was
subjected to the judgment of four trained panel members who classify the samples in a scale according to olfac-tory sensations, gustaolfac-tory–retronasal sensations, and final olfactory–gustatory sensations. For olfactory sensations, the following attributes were measured: olive fruitiness (0–7); other fruits (0–3); green (grass/leaves) (0–2); other positive sensations (0–3); and harmony (0–20). Concerning gustatory–retronasal sensations were evaluated the olive fruitiness (0–10); sweet (0–4); bitter (0–3); pungent (0–3); green (grass/leaves) (0–2); other positive sensations (0–3); and harmony (0–20). A final olfactory–gustatory sensation for each sample was also considered, conjugating all the organoleptic sensations, pointing out the complexity (0–10) and persistence (0–10). It should be remarked that harmony increases when the attributes are balanced, and complexity increases with the number and intensity of aromas and fla-vors. Although COI guidelines recommend the use of 8–12 tasters for each analysis, it was decided to use only four trained panelists (three from Portugal and one from Spain), which are quite familiar with the classification of positive and negative olive oil organoleptic attributes and usually are jury members in olive oil classification in competitions. With this approach, it was intended to guarantee a more consistent sensory classification.
Sample preparation and E‑tongue analysis
All samples were electrochemically analyzed in the same day, which turned out in different storage times, namely 1 year or less. The olive oil extraction procedure, using
hydroethanolic solutions (H2O:EtOH, 80:20 v/v), was
pre-viously reported by Dias et al. [3]. One independent
sam-ple was collected and extracted from each bottle of each monovarietal EVOO and analyzed in duplicate (totalizing 22 olive oils, 44 bottles, and 88 analyses). The E-tongue included two print-screen potentiometric devices, con-taining different cross-sensitivity membranes as
chemi-cal sensors [3]: four lipidic additives (octadecylamine,
oleyl alcohol, methyltrioctylammonium chloride, and oleic acid from Fluka; corresponding to approximately
3 %), five plasticizers (bis(1-butylpentyl) adipate, dibutyl sebacate, 2-nitrophenyl-octylether, tris(2-ethylhexyl)phos-phate, and dioctyl phenylphosphonate from Fluka; repre-senting around 65 %) and high molecular weight polyvi-nyl chloride (PVC); near 32 %). The type of sensors and polymeric membrane compositions (relative percentage of additive, plasticizer, and PVC) were selected based on
a previous work [33] taking into account their satisfactory
signal stability over time (%RSD < 5 %) and repeatability (0.5 % < %RSD < 15 %) for basic standard taste com-pounds (e.g., sweet, acid, bitter, salty, and umami). Further
details about each sensor are described by Sousa et al. [34].
Also, lipid polymeric membranes were used since they may interact with tastant substances through electrostatic
or hydrophobic interactions [35]. Each sensor was
identi-fied with a code with a letter S (for sensor) followed by the number of the array (1 or 2) and the number of the mem-brane (1–20, corresponding to different combinations of plasticizer and additive used).
Data fusion
Three data fusion approaches are usually applied. The main differences rely on the abstraction levels, i.e., the way how data originated from several analytical techniques or different sources, can be merged, and form a consistent concatenated data matrix. In low-level of abstraction data fusion, the data from all sources are simply concatenated before model development, resulting in a data matrix with a dimension of number of samples × total number of vari-ables from all sources. This approach may improve model classification performance namely multisensor data fusion
[19]. In mid-level fusion, a feature extraction is applied to
each data source followed by merging the extracted fea-tures. For high level of abstraction, data from each source are initially analyzed, and a model is derived, separately, from each data source, being the model classification
out-puts then merged [26, 28, 36]. In this work, a low-level data
fusion was adopted.
Statistical analysis
Different chemometric methods were applied. First, sen-sory data were subjected to an analysis of variance (two-way ANOVA) to infer about the existence of a significant effect of production year or olive cultivar. If a significant effect was found, the Tukey’s post hoc multiple comparison test was further employed.
Four linear multivariate methods were explored to estab-lish monovarietal EVOO classification models using inde-pendent predictors among the E-tongue potentiometric signals recorded and/or the sensory attributes assessed by the trained panelists, namely linear discriminant analysis
(LDA) and partial least squares discriminant analysis (PLS-DA), as the most common linear techniques applied in clas-sification studies; sparse partial least squares discriminant analysis (SPLS-DA), as a PLS improved methodology, which allows to obtain a simplified PLS model by applying a feature variable selection; and finally, a LDA approach coupled with meta-heuristic simulated annealing algorithm (LDA-SA), which uses a variable selection step.
The LDA is a supervised technique used as a starting point in classification studies since the model obtained is represented by linear functions of the original variables. So, the importance of the variables in the classification
pro-cess can be easily perceived [37]. The PLS-DA is a more
complex supervised classification technique and should be used if a deeper mathematical treatment is required. This technique is usually applied when p explanatory variables show high multicollinearity and/or are greater than the total number of samples/observations (N). The PLS procedure implies a dimension reduction step by forming new orthog-onal variables (linear combinations of the original vari-ables, namely principal component (PCs) functions, which coefficients involve information about the variances and covariances of both the explanatory and response variables. The model obtained is complex since new variables are
represented in a new dimensional space [37]. The
SPLS-DA is based in the sparsity principle [38], on the basis that
not all explanatory variables are needed to the fit model, i.e., meaning that a subset of original variables could be mainly responsible for determining the response. There-fore, SPLS-DA also has the ability to perform feature selec-tion keeping the structure of PLS-DA model. Considering that some variables may enhance noise effects in PLS-DA prediction, the SPLS-DA technique imposes sparsity on the dimension reduction step of PLS by selecting variables
and increasing the predictive power of the model [37]. As
in PLS technique, SPLS uses a data matrix computed by principal components and not the original information of explanatory and response variables. A major disadvantage of LDA compared with the other two other multivariate techniques is the lack of robustness regarding the existence of collinear variables. To overcome this drawback, LDA was also implemented coupled with a meta-heuristic simu-lated annealing (SA) variable selection algorithm, enabling the selection of the best subsets of independent predictors among the E-tongue potentiometric signals recorded and/ or the sensory sensations assessed by the trained panelists. This algorithm pursues to select the optimal conditions based on the annealing physic process, which mimics the slow controlled cooling process of a heated material in order that the material can reach the most regular possible crystal lattice configuration (free from defects at a mini-mum energy crystalline state). It is classified as a meta-heu-ristic algorithm, since it is able to identify a subset of the
original independent variables that correspond to a global optimum for a given approximation criterion, selected within a large search space of other possible subsets of
variables [39]. This means that for optimizing a system
by selecting the best subset of k variables considering this type of thermodynamic behavior, in each iteration, values of two solutions (the current subset of k variables and the new subset to be tested, also with k variables) are compared according to a criterion that measures the quality of those subsets of variables (for instance, accuracy of a classifica-tion model). The new soluclassifica-tion is randomly selected in the neighborhood of the current solution and tested according to the simulated annealing rules, becoming the new solu-tion if the criterion has best values than the original. The algorithm continues the search for new solutions until it reaches the maximum number of interactions established at the beginning of procedure.
To evaluate the classification models’ stability and qual-ity, a leave-one-out cross-validation (LOO-CV) procedure was applied. Although this methodology may be consid-ered an overoptimistic procedure, the use of a more robust cross-validation approach (e.g., based on an external vali-dation set) was not possible due to the large number of groups considered (eight monovarietal EVOOs) and the reduction in the number of independent olive oil samples for each group (between 2 and 4). Indeed, when the num-ber of samples is small, LOO-CV procedure has proven to be an adequate procedure to test model prediction capabil-ity and to minimize the risk of overoptimistic predictive
model performance [3, 25, 40].
So, as a first step, the discriminative capability of the olfactory–gustatory sensations was tested, being the qual-ity of the established models (LDA, PLS-DA, SPLS-DA, and LDA-SA) were evaluated based on the LOO-CV pro-cedure. Secondly, the same procedure was applied to the potentiometric signal profiles, and as before, their classifi-cation capabilities were also evaluated using the LOO-CV methodology. To the best model, the spatial distribution of the single-cultivar EVOO original groups was tentatively correlated with the taste sensory attributes evaluated by the trained panelists, by means of the linear Pearson correla-tion coefficient (R-Pearson), for inferring about the practi-cal potential of the E-tongue in mimicking the human taste sensory perception of the monovarietal EVOO analyzed.
Finally, a low-level of abstraction data fusion approach was applied, combining sensor signal profiles and sensory descriptors of the olive oils, to evaluate synergies in the information contained in both data sets. Data set size of each source is of similar magnitude (44 × 40 and 44 × 14; for E-tongue and sensory trained panel analysis, respec-tively) minimizing the possibility of the largest data set
collapsed the other [28, 41, 42]. To normalize the weight
classification models, variable scaling and centering pro-cedures were applied. The main goal of this approach was to improve Spanish monovarietal EVOO discrimination according to olive cultivar.
All statistical analyses were performed using the caret
[37, 43], Subselect [39, 44], pls [37, 45], spls [37, 46], and
MASS [47] packages of the open-source statistical
pro-gram R (version 2.15.1). Results and discussion
Sensory data evaluation
Each monovarietal EVOO sample was evaluated by four trained sensory panelists and classified according to 14 organoleptic descriptors including five olfactory sensations, seven gustatory–retronasal sensations, and two final olfac-tory–gustatory sensations, according to the IOC directives
[31, 32]. The two-way ANOVA showed that some olive oil
sensory descriptors were significantly affected by the pro-duction year (olfactory sensations: olive fruitiness, green, and harmony; gustatory–retronasal sensations: bitter, pun-gent, green, and harmony; final olfactory–gustatory sensa-tions: persistence; p values <0.0154) and olive oil cultivar (olfactory sensations: olive fruitiness and harmony; gusta-tory–retronasal sensations: sweet, pungent, other positive sensations, and harmony; final olfactory–gustatory sensa-tions: persistence; p values <0.0357), but the interaction effects were not statistically significant (p value >0.1050) at a 5 % significance level. The differences found between production years may be related to the different edapho-climatic conditions. However, in this work, olive oils were used, regardless of their production year, trying to evaluate the potentiality of the data fusion approach. The significant effect of olive cultivar in some olive oil sensory sensa-tions may be attributed to their different physicochemical composition, which affects the organoleptic characteristics
such as pungency, bitterness, and astringency. In Fig. 1,
a comparison of mean sensory data from the quantitative descriptive analysis (QDA), for the eight single-cultivar EVOO evaluated, is shown. The radar plot, for each sen-sory descriptor, also shows which olive cultivar means are significantly different at a 5 % level, according to the post hoc Tukey’s multiple comparison test. Globally, it can be stated that EVOO from cvs. Redondilla and Cornicabra are those that show more differences within olfactory sensa-tions. Based on the gustatory–retronasal sensations, it was found that EVOOs from cvs. Manzanilla, Redondilla, and Royuela were those that showed the greater differences.
Finally, the capability of the sensory descriptors to dis-criminate the different single-cultivar EVOO was assessed by means of the supervised linear techniques: LDA,
PLS-DA, SPLA-DA, and LDA-SA. Table 2 shows the
accuracy obtained for all these multivariate techniques. The LDA and PLS-DA results were obtained considering all the independent variables in the model. For PLS-DA and SPLS-DA approaches, the number of PCs in the model was select from a search process with a maximum number of 14 possible PCs [equal to the number of PCs defined by prin-cipal component analysis (PCA)]. Also, the SPLS-DA per-formed feature variable selection within the PCs allowing to increase the predictive power of the model. The LDA-SA uses LDA-SA algorithm to identify a subset of the original independent variables. To establish all the multivariate clas-sification models, the global optimum criterion used was the accuracy (i.e., the level of correct classifications using the LOO-CV procedure also called sensitivity).
The best result concerning single-cultivar EVOO classi-fication based on olive oil sensory attributes (greater accu-racy percentage) was obtained from the LDA-SA technique (57 % of correct classifications for LOO-CV, according to olive cultivar). The worst performances were achieved by LDA and PLS-DA methods (39 % of correct classifica-tions). Since among the organoleptic attributes no
multicol-linearity issues were found (R2-Pearson < 0.48), the poor
results of LDA and PLS-DA techniques may be due to the presence of variables that increase noise effects affecting their prediction capabilities, being this problem slightly reduced, but not completely eliminated, with the variable selection procedure of SPLS-DA technique.
The SA algorithm used with the LDA technique, allowed selecting a set of seven sensory attributes (two olfactory sensations: olive fruitiness and other fruits; four gustatory–retronasal sensations: sweet, bitter, and pun-gency; and one final olfactory–gustatory sensation: com-plexity) from the 14 available organoleptic sensations, resulting in a linear discriminant model with seven func-tions (100 % of explained variance). The model obtained enabled a correct classification percentage of only 84 and
57 % for the original data group (Fig. 2, for the first three
discriminant functions) and for the LOO-CV procedure. So, it is clear that, although the above-mentioned sensory descriptors may contribute to EVOO classification accord-ing to olive cultivar, they cannot be used for discrimina-tion issues, and alone, these attributes cannot be applied to guarantee correct monovarietal EVOO labeling. Indeed, except for EVOO from cvs. Redondilla and Zorzal, all other single-cultivar EVOOs present misclassification problems. Nevertheless, it should be noticed that from the sensory descriptors assessed by the panelists, gustatory– retronasal sensations were those that most contributed for olive cultivar recognition. In conclusion, the known capa-bility of human sensory evaluation to assess organoleptic characteristics of virgin olive oil used for quality classifica-tion seems to be not sufficient to discriminate high-quality
Other.positive.sensations ARBO ab ARR ab COR ab FRA a MAN ab RED ab ROY ab ZOR b 0 0.5 1 1.5 2 2.5 3 Harmony ARBO bc ARR abc COR bc FRA ab MAN c RED a ROY bc ZOR abc 0 5 10 15 20 Complexity ARBO a ARR a COR a FRA a MAN a RED a ROY a ZOR a 0 2 4 6 8 10 Persistence ARBO a ARR a COR a FRA a MAN b RED a ROY a ZOR ab 0 2 4 6 8 10 Olfactory-Gustatory sensations Olive.Fruitiness ARBO a ARR a COR a FRA a MAN a RED a ROY a ZOR a 0 2 4 6 8 10 Sweet ARBO ab ARR ab COR ab FRA bc MAN a RED a ROY c ZOR a 0 1 2 3 4 Bitter ARBO ab ARR ab COR b FRA ab MAN b RED ab ROY a ZOR ab 0 0.5 1 1.5 2 2.5 3 Pungent ARBO a ARR a COR a FRA a MAN b RED ab ROY a ZOR a 0 0.5 1 1.5 2 2.5 3 Green ARBO a ARR a COR a FRA a MAN a RED a ROY a ZOR a 0 0.5 1 1.5 2 Olive.Fruitiness ARBO ab ARR b COR b FRA ab MAN b RED a ROY ab ZOR ab 0 1 2 3 4 5 6 7 Other.Fruits ARBO a ARR a COR a FRA a MAN a RED a ROY a ZOR a 0 0.5 1 1.5 2 2.5 3 Green ARBO a ARR a COR a FRA a MAN a RED a ROY a ZOR a 0 0.5 1 1.5 2 Ohter.positive.sensations ARBO a ARR a COR a FRA a MAN a RED a ROY a ZOR a 0 0.5 1 1.5 2 2.5 3 Harmony ARBO ab ARR bc COR c FRA ab MAN abc RED a ROY ab ZOR ab 0 5 10 15 20
Olfactory sensations Gustatory-retronasal
sensations Gustatory-retronasal sensations
Fig. 1 Radar plot of the olive oil sensory descriptors assessed by the trained panelists for each monovarietal EVOO (ARB: cv. Arbosana; ARR: cv. Arroniz; COR: cv. Cornicabra; FRA: cv. Frantoio; MAN: cv. Manzanilla; RED: cv. Redondilla; ROY: cv. Royuela; and ZOR:
cv. Zorzal). Different letters for the same sensory descriptor point out significant statistical differences identified by Tukey’s post hoc multi-ple comparison test (p < 0.05)
monovarietal EVOO according to olive cultivar. Neverthe-less, its potential exists and could be complemented by an artificial electronic sensory device like the electronic tongue.
E‑tongue signals analysis
On the whole, 88 assays were carried out (22 samples of eight EVOO × 2 bottles × 1 extraction of each bottle × 2 analysis), each providing 40 potentiometric signals, vary-ing from -0.03 V to +0.25 V (20 different sensor mem-branes used in duplicate: S1:1–S1:20 and S2:1–S2:20, as
previously discussed [3, 40]). Since the voltage signals,
recorded during the analysis of the hydroethanolic extracts of the monovarietal EVOO studied, are similar for all sen-sors and the respective range of variation is relatively
nar-row, as shown in Fig. 3 for two EVOOs (e.g., ARR and
MAN), there was no need of data scaling. In fact, for the duplicate analysis of the hydroethanolic extract of each of the 44 olive oil bottles, the %RSD values of the signals recorded with each one of 40 lipid membranes varied from 0.25 to 1.8 %; for the analysis of the two bottles collected for each monovarietal EVOO, the %RSD ranged from 5.6 to 8.7 %; and finally, for the analysis of all bottles col-lected, the %RSD varied from 4.6 to 6.9 %. These results showed satisfactory accuracy of the E-tongue analysis.
The results from the application of the four linear mul-tivariate techniques using the E-tongue potentiometric signal profiles to discriminate the different single-cultivar
EVOOs are presented in Table 2. As in the previous
sec-tion, the comparison between the four techniques was made based on the accuracies for LOO-CV procedure. In this case, the optimum number of PCs for PLS-DA and SPLS-DA models was selected within 38 possible PCs (number of PCs defined by PCA). Again, the best perfor-mance was obtained from the LDA-SA technique (70 % of correct classifications for LOO-CV) compared with the three other linear multivariate techniques (LDA, PLS-DA, and SPLS-DA: ≤27 % of correct classifications). Indeed, the SA variable selection procedure allowed choosing the most relevant signals for establishing the best predictive LDA model, minimizing the risk of using high collinear variables. So, the resulting LDA-SA classification model had three significant discriminant functions, explaining 98.86 % of the original data variability. The model was based on 26 non-collinear sensor signals (S1:2, S1:4, S1:5, S1:7, S1:9–S1:13, S1:15; S1:17, S1:18, S2:1–S2:3, S2:5– S2:7, S2:9–S2:11, S2:13, S2:14; S2:17–S2:19), allowing 100 and 70 % of correct classifications for original data
(Fig. 4) and LOO-CV procedure, respectively. The SA
Table 2 Discriminative capability of the sensory descriptors (olfac-tory–gustatory sensations), potentiometric signal profiles, and low-level of abstraction data fusion approach was applied, combining sensor signal profiles and sensory descriptors of the olive oils, toward single-cultivar EVOO differentiation, based on the accuracy values in the LOO-CV procedure
a Accuracy = % of correct classifications
Data matrix Accuracy for LOO-CV (%)a
LDA PLS-DA SPLS-DA LDA-SA
Sensory attributes 39 39 45 57
No of selected sensory attributes
14 14 7 7
No of principal components – 8 4 –
E-tongue sensor signal profiles 23 27 27 70 No of selected E-tongue signals 40 40 29 26 No of principal components – 24 1 –
Low-level data fusion 7 41 48 100
No of selected sensory attributes 14 14 4 9 No of selected E-tongue signals 40 40 8 18 No of principal components – 10 5 – −5.0 −2.5 0.0 2.5 −6 −3 0 3 6
First discriminant function (65.0%)
Second discriminant function (18.3%)
ARBO ARR COR FRA MAN RED ROY ZOR
Fig. 2 Spanish monovarietal EVOO classification performance for the sensory attributes data: first and second LDA functions based on the best subset of sensory descriptors (seven olfactory, gustatory–ret-ronasal, and final olfactory–gustatory sensations), assessed by trained panelists, selected using the SA meta-heuristic algorithm (ARB: cv. Arbosana; ARR: cv. Arroniz; COR: cv. Cornicabra; FRA: cv. Fran-toio; MAN: cv. Manzanilla; RED: cv. Redondilla; ROY: cv. Royuela; and ZOR: cv. Zorzal)
algorithm selected seven repeated sensors, with the same membrane composition. The inclusion of repeated sensors could be justified by their slight different physical proper-ties due to the possible inhomogeneous membrane surface, which could be attributed to the drop-by-drop membrane
preparation technique [40]. On the other hand, the inclusion
of repeated sensors in a multivariate model may increase its
qualitative and/or quantitative performance [48].
It should be noticed that 65 % of the E-tongue sensors (17 of the 26) used in the classification model are the same
as reported by Dias et al. [3] for successfully
discriminat-ing three Spanish monovarietal EVOOs (cvs. Arbequina, Hojiblanca, and Picual), which were also produced in the same geographical region as those analyzed in the present work. Although the predictive LDA-SA model sensitivity was not very satisfactory (only EVOO samples from cvs. Arbosana, Arroniz, and Royuela were not misclassified), a clear increase was obtained compared to that of the LDA-SA model established using sensory descriptors evaluated
by the trained panelists. Also, as already pointed out [3],
the capability of the potentiometric E-tongue to correctly classify single-cultivar EVOO according to olive culti-var may be attributed to the nature and level of the polar compounds present in the EVOO extracts analyzed, which could be related to sensory olive oil attributes such as bit-terness, astringency, and pungency. Indeed, this evidence was further demonstrated in the present work, since the spatial distribution of the EVOO group centroids (mean value of the discriminant scores for a given category) of the first linear discriminant function (cvs. centroids order: Royuela < Arroniz < Frantoio < Zorzal < Arbosana < Cor-nicabra < Redondilla < Manzanilla), which explained more than 95 % of the potentiometric E-tongue signals variability, could be statistically correlated with gusta-tory–retronasal descriptors assessed by the trained pan-elists: bitter (R-Pearson = −0.83; p value = 0.0110), pun-gent (R-Pearson = −0.91; p value = 0.0017), and green Fig. 3 E-tongue potentiometric
signal profiles of the hydroetha-nolic extracts of two monovari-etal EVOO (ARR: cv. Arroniz; MAN: cv. Manzanilla) 0.10 0.15 0.20 0.25 Number of sensor P otential (V) S1:1 S1:2 S1:3 S1:4 S1:5 S1:6 S1:7 S1:8 S1:9 S1:10 S1:11 S1:12 S1:13 S1:14 S1:15 S1:16 S1:17 S1:18 S1:19 S1:20 S2:1 S2:2 S2:3 S2:4 S2:5 S2:6 S2:7 S2:8 S2:9 S2:10 S2:11 S2:12 S2:13 S2:14 S2:15 S2:16 S2:17 S2:18 S2:19 S2:20 ARR MAN −5 0 5 −20 0 20
First discriminant function (95.2%)
Second discriminant function (2.2%)
ARBO ARR COR FRA MAN RED ROY ZOR
Fig. 4 Spanish monovarietal EVOO classification performance for the sensor signals data: first and second LDA functions based on the best subset of E-tongue potentiometric signals (26 sensors) selected using the SA meta-heuristic algorithm (ARB: cv. Arbosana; ARR: cv. Arroniz; COR: cv. Cornicabra; FRA: cv. Frantoio; MAN: cv. Manza-nilla; RED: cv. Redondilla; ROY: cv. Royuela; and ZOR: cv. Zorzal)
(R-Pearson = −0.80; p value = 0.0180) sensations. These results strengthen the idea that the potentiometric E-tongue can mimic olive oil gustatory–retronasal human sensations, which was shown for the first time using sensory sensations data. Then again, this behavior was expected since this type of lipid polymeric membranes could recognize and assess standard solutions of basic tastant substances such as acid,
bitter, salty, sweet, and umami [33].
Low‑level data fusion of sensory attributes and E‑tongue signals
The results previously obtained using sensory attributes or E-tongue signals individually showed that both meth-odologies could contribute to monovarietal EVOO dis-crimination by olive cultivar although with unsatisfactory performances. So, the simultaneous use of both human and artificial evaluations was further explored trying to demon-strate their complementary performances. For that, a low-level fusion approach was implemented, merging the 40 potentiometric signals recorded during the electrochemi-cal analysis of the monovarietal EVOO hydroethanolic extracts, with the 14 sensory attributes evaluated by the trained panelists. A single matrix containing 44 samples by 54 variables was obtained.
The accuracy of the low-level fusion approach of the sensory attributes and E-tongue signals to discriminate the different single-cultivar EVOO obtained by applying each of the four linear multivariate techniques is presented in
Table 2. In this study, the number of PCs established for
PLS-DA and SPLS-DA techniques was set based on a search procedure within 42 possible PCs (number of PCs defined by PCA). Once again, the results confirmed that the LDA-SA technique established the most powerful predic-tive model (100 % of correct classification for LOO-CV), followed by SPLS-DA, although with a quite lower classifi-cation predictive capability (only 48 % of correct classified samples), showing nevertheless that the variable selection procedures are advantageous, allowing to obtain simpler models with better predictive capabilities. It should be remarked that the low-level data fusion was only an effec-tive classification procedure when combined with LDA-SA technique. Indeed, simple human sensory evaluation cou-pled with SPLS-DA or LDA-SA, or E-tongue data coucou-pled with LDA-SA showed similar or higher classification per-formances (ranging from 45 to 70 % of correct
classifica-tions as shown in Table 2).
Regarding the best linear multivariate technique, a LDA-SA model with two discriminant functions, explaining 100 % of the original data variability (99.91 and 0.03 %, respectively), was established based on signal profiles of 18 potentiometric E-tongue sensors (S1:1, S1:3, S1:6–S1:8, S1:12, S1:13, S1:19, S1:20, S2:9, S2:10, S2:12–S2:15,
S2:17, S2:18, and S2:20) and nine sensory attributes (olfac-tory sensations: other fruits and other positive sensations; gustatory–retronasal sensations: olive fruitiness, bitter, pun-gent, green, other positive sensations, and harmony; final olfactory–gustatory sensations: complexity), selected by the SA algorithm. The SA variable selection algorithm ena-bled the inclusion of only three repeated sensors, with the same membrane composition, in the discrimination model, which could improve the global model of the multivariate
statistical performance [48]. Although the second
discrimi-nant function LDA-SA obtained model was not significant in the classification of groups of different varieties of olive oil, it was used to establish a two-dimensional discriminant graphic in order to better visualize the groups separation.
The proposed model allowed a 100 % correct
classifi-cation of the original data (Fig. 5), as well as 100 %
pre-dictive correct classification for the LOO-CV procedure, as previously mentioned. This result shows that the E-tongue device fused with sensory attributes could be used to cor-rectly classify monovarietal EVOO according to correct olive cultivar and so could be used as practical tool to guarantee the correct labeling of single-cultivar EVOO for cvs. Arbosana, Arroniz, Cornicabra, Frantoio, Manzanilla, Redondilla, or Royuela.
So, this work showed the potential use of electrochemi-cal sensors to evaluate olive oil quality. Indeed, electronic
−4 0 4
−200 0 200
First discriminant function (99.91%)
Second discriminant function (0.03%)
ARBO ARR COR FRA MAN RED ROY ZOR
Fig. 5 Spanish monovarietal EVOO classification performance for the E-tongue data using a low-level data fusion approach: first and second LDA functions based on the best subset E-tongue (18 sen-sors) and sensory descriptors (nine olfactory, gustatory–retronasal, and final olfactory–gustatory sensations), selected using the SA meta-heuristic algorithm (ARB: cv. Arbosana; ARR: cv. Arroniz; COR: cv. Cornicabra; FRA: cv. Frantoio; MAN: cv. Manzanilla; RED: cv. Redondilla; ROY: cv. Royuela; and ZOR: cv. Zorzal)
noses (E-noses) and electronic tongues (E-tongues), indi-vidually or combined, have been proposed in the last dec-ade for olive oil characterization using different multivari-ate statistical techniques. For instance, an E-nose has been successfully used for the discrimination of different
Medi-terranean single-cultivar EVOO [49]. More recently, our
research team has shown that an E-tongue could be applied to successfully discriminate Portuguese (100 % of correct classification) or Spanish (97.5–100 % of correct classifica-tions) monovarietal EVOO according to the olive cultivar, depending on the number of olive cultivars evaluated (3 or
11) [3, 20].
Regarding monovarietal olive oil classification accord-ing to olive cultivar, other approaches have been described in the literature, reporting in some cases slightly less accu-rate correct classification performances (ranging from 70 to 100 %) depending on the analytical technique used (e.g., NMR, NIR, and MIR spectroscopy; PTR-MS; HS-SPME-GC/MS; SNP-based PCR–RFLP capillary electrophoresis; SNP-based CAPS assays), the number of olive cultivars under study (3, 5, or 10 cultivars) and chemometric tool used (e.g., LDA PLS-DA and canonical discriminant
analy-sis) [13, 14, 22, 50–52].
Conclusions
Previously it was proven that a potentiometric E-tongue could be used to classify monovarietal EVOO. However, the successful discrimination performance was obtained when few single-cultivar EVOOs were evaluated. In this work, it was demonstrated that the proposed approach could be extended to a greater number single-cultivar olive oils when a low-level data fusion approach was adopted, merging E-tongue and human sensory analysis. The pro-posed strategy enables a full predictive LOO-CV discrimi-nation of olive oil samples from eight monovarietal EVOOs (100 % correct classifications), produced in two different years. Moreover, for the first time, it was shown that the E-tongue could mimic human gustatory–retronasal sensa-tions of EVOO. The successful classification was achieved using a LDA-SA model, based on the most informative potentiometric signals and sensory descriptors, chosen using the SA variable selection algorithm. This methodol-ogy proved to be more powerful than the two most common linear supervised multivariate techniques, LDA and PLS-DA, and even SPLS-PLS-DA, a PLS improved methodology that allows feature variable selection. So, it could be concluded that, from a practical point of view, E-tongue and sensory analysis performed by trained panelists are complementary techniques, contributing to improve the classification rate of monovarietal EVOO according to olive cultivar. Also, it is shown that the traditional use of olive oil panel sensory
analysis could be extended to the classification of mono-varietal EVOO according to olive cultivar if a complemen-tary approach was implemented by fusing human (sensory panel) and artificial (electronic tongue) sensory assessment. So, the use of this type of tool may indeed allow guarantee-ing sguarantee-ingle-cultivar EVOO correct labelguarantee-ing, minimizguarantee-ing the risk of frauds of this high-value product.
Acknowledgments This work was co-financed by FCT/MEC and FEDER under Program PT2020 (Project UID/EQU/50020/2013); by Fundação para a Ciência e Tecnologia under the strategic funding of UID/BIO/04469/2013 unit; and by Project POCTEP through Project RED/AGROTEC—Experimentation network and transfer for devel-opment of agricultural and agro industrial sectors between Spain and Portugal.
Compliance with ethical standards
Conflict of interest The authors declare that they have no conflict of interest.
Compliance with Ethics Requirements This article does not con-tain any studies with human or animal subjects.
References
1. Sinelli N, Cerretani L, Di Egidio V, Bendini A, Casiraghi E (2010) Application of near (NIR) infrared and mid (MIR) infra-red spectroscopy as a rapid tool to classify extra virgin olive oil on the basis of fruity attribute intensity. Food Res Int 43:369–375 2. Garcia R, Martins N, Cabrita MJ (2013) Putative markers of
adulteration of extra virgin olive oil with refined olive oil: pros-pects and limitations. Food Res Int 54:2039–2044
3. Dias LG, Fernandes A, Veloso ACA, Machado AASC, Pereira JA, Peres AM (2014) Single-cultivar extra virgin olive oil clas-sification using a potentiometric electronic tongue. Food Chem 160:321–329
4. Lerma-García MJ, Simó-Alfonso EF, Bendini A, Cerretan L (2009) Metal oxide semiconductor sensors for monitoring of oxi-dative status evolution and sensory analysis of virgin olive oils with different phenolic content. Food Chem 117:608–614 5. Rotondi A, Beghè D, Fabbri A, Ganino T (2011) Olive oil
trace-ability by means of chemical and sensory analyses: a comparison with SSR biomolecular profiles. Food Chem 129:1825–1831 6. Lauri I, Pagano B, Malmendal A, Sacchi R, Novellino E,
Ran-dazzo A (2013) Application of “magnetic tongue” to the sensory evaluation of extra virgin olive oil. Food Chem 140:692–699 7. Carrasco-Pancorbo A, Gómez-Caravaca AM, Cerretani L,
Bend-ini A, Segura-Carretero A, Fernández-Gutiérrez A (2006) Rapid quantification of the phenolic fraction of Spanish virgin olive oils by capillary electrophoresis with UV detection. J Agric Food Chem 54:7984–7991
8. Matos LC, Cunha SC, Amaral JS, Pereira JA, Andrade PB, Seabra RM, Oliveira BPP (2007) Chemometric characterization of three varietal olive oils (cvs. Cobrançosa, Madural and Ver-deal Transmontana) extracted from olives with different matura-tion indices. Food Chem 102:406–414
9. Bakhouche A, Lozano-Sánchez J, Beltrán-Debón R, Joven J, Segura-Carretero A, Fernández-Gutiérrez A (2013) Phenolic characterization and geographical classification of commercial Arbequina extra-virgin olive oils produced in southern Catalonia. Food Res Int 50:401–408
10. Karabagias I, Michos Ch, Badeka A, Kontakos S, Stratis I, Kon-tominas MG (2013) Classification of Western Greek virgin olive oils according to geographical origin based on chromatographic, spectroscopic, conventional and chemometric analyses. Food Res Int 54:1950–1958
11. Longobardi F, Ventrella A, Napoli C, Humpfer E, Schütz B, Schäfer H, Kontominas MG, Sacco A (2012) Classification of olive oils according to geographical origin by using 1H NMR fingerprint-ing combined with multivariate analysis. Food Chem 130:177–183 12. Romero C, Brenes M (2012) Analysis of total contents of
hydroxytyrosol and tyrosol in olive oils. J Agric Food Chem 60:9017–9022
13. Ruiz-Samblás C, Tres A, Koot A, van Ruth SM, González-Cas-ado A, Cuadros-Rodríguez L (2012) Proton tranfer reaction-mass spectrometry volatile organic compound fingerprint for monovarietal extra virgin olive oil identification. Food Chem 134:589–596
14. Bazakos C, Dulger AO, Uncu AT, Spaniolas S, Spano T, Kalaitzis P (2012) A SNP-based PCR–RFLP capillary electrophoresis analysis for the identification of the varietal origin of olive oils. Food Chem 134:2411–2418
15. Dais P, Hatzakis E (2013) Quality assessment and authentication of virgin olive oil by NMR spectroscopy: a critical review. Anal Chim Acta 765:1–27
16. Nunes CA (2013) Vibrational spectroscopy and chemometrics to assess authenticity, adulteration and intrinsic quality parameters of edible oils and fats. Food Res Int 60:255–261
17. Apetrei IM, Apetrei C (2013) Voltammetric e-tongue for the quantification of total polyphenol content in olive oils. Food Res Int 54:2075–2082
18. Escuderos ME, Sánchez S, Jiménez A (2011) Quartz Crystal Microbalance (QCM) sensor arrays selection for olive oil sen-sory evaluation. Food Chem 124:857–862
19. Haddi Z, Alami H, El Bari N, Tounsi M, Barhoumi H, Maaref A, Jaffrezic-Renault N, Bouchikhi B (2013) Electronic nose and tongue combination for improved classification of Moroccan vir-gin olive oil profiles. Food Res Int 54:1488–1498
20. Peres AM, Veloso ACA, Pereira JA, Dias LG (2014) Electro-chemical multi-sensors device coupled with heuristic or meta-heuristic selection algorithms for single-cultivar olive oil classifi-cation. Proced Eng 87:192–195
21. Casale M, Casolino C, Oliveri P, Forina M (2010) The potential of coupling information using three analytical techniques for identifying the geographical origin of Liguria extra virgin olive oil. Food Chem 118:163–170
22. Casale M, Sinelli N, Oliveri P, Di Egidio V, Lanteri S (2010) Chemometrical strategies for feature selection and data compres-sion applied to NIR and MIR spectra of extra virgin olive oils for cultivar identification. Talanta 80:1832–1837
23. Casale M, Oliveri P, Casolino C, Sinelli N, Zunin P, Armanino C, Forina M, Lanteri S (2012) Characterisation of PDO olive oil Chianti Classico by non-selective (UV–visible, NIR and MIR spectroscopy) and selective (fatty acid composition) analytical techniques. Anal Chim Acta 712:56–63
24. Pizarro C, Rodríguez-Tecedor S, Pérez-del-Notario N, Esteban-Díez I, González-Sáiz JM (2013) Classification of Spanish extra virgin olive oils by data fusion of visible spectroscopic finger-prints and chemical descriptors. Food Chem 138:915–922 25. Gutiérrez JM, Haddi Z, Amari A, Bouchikhi B, Mimendia A,
Cetó X, del Valle M (2013) Hybrid electronic tongue based on multisensor data fusion for discrimination of beers. Sens Actua-tors B 177:989–996
26. Vera L, Aceña L, Guasch J, Boqué R, Mestres M, Busto O (2011) Characterization and classification of the aroma of beer samples by means of an MS e-nose and chemometric tools. Anal Bioanal Chem 399:2073–2081
27. Bruwer M-J, MacGregor JF, Bourg WM Jr (2007) Fusion of sensory and mechanical testing data to define measurements of snack food texture. Food Qual Prefer 18:890–900
28. Haddi Z, Mabrouk S, Bougrini M, Tahri K, Sghaier K, Bar-houmi H, El Bari N, Maaref A, Jaffrezic-Renault N, Bouchikhi B (2014) E-nose and e-tongue combination for improved recogni-tion of fruit juice samples. Food Chem 150:246–253
29. Banerjee R, Modak A, Mondal S, Tudu B, Bandyopadhyay R, Bhattacharyya N (2013) Fusion of electronic nose and tongue response using fuzzy based approach for black tea classification. Procedia Technol 10:615–622
30. Apetrei C, Apetrei IM, Villanueva S, de Saja JA, Gutierrez-Rosales F, Rodriguez-Mendez ML (2010) Combination of an e-nose, an e-tongue and an e-eye for the characterisation of olive oils with different degree of bitterness. Anal Chim Acta 663:91–97
31. International Olive Council (2013) Sensory analysis of olive oil—method for the organoleptic assessment of virgin olive oil. COI/T.20/Doc. No. 15/Rev. 6 November 2013. http://www.inter-nationaloliveoil.org/
32. International Olive Council (2014) IOC Mario Solinas quality award—rules of the international competition for extra virgin olive oils. T.30/Doc. No. 17 June 2014. http://www.internation-aloliveoil.org/
33. Dias LG, Peres AM, Veloso ACA, Reis FS, Vilas Boas M, Machado AASC (2009) An electronic tongue taste evaluation: identification goat milk adulterations with bovine milk. Sens Actuators B 136:209–217
34. Sousa MEBC, Dias LG, Veloso ACA, Estevinho L, Peres AM, Machado AASC (2014) Practical procedure for discriminating monofloral honeys with a broad pollen profile variability using an electronic tongue. Talanta 128:284–292
35. Kobayashi Y, Habara M, Ikezazki H, Chen R, Naito Y, Toko K (2010) Advanced taste sensors based on artificial lipids with global selectivity to basic taste qualities and high correlation to sensory scores. Sensors 10:3411–3443
36. Liu Y, Brown SD (2004) Wavelet multiscale regression from the perspective of data fusion: new conceptual approaches. Anal Bioanal Chem 380:445–452
37. Kuhn M, Johnson K (2013) Applied predictive modeling, fea-tures. Springer, 17 May 2013
38. Chun H, Keles¸ S (2010) Sparse partial least squares regression for simultaneous dimension reduction and variable selection. J R Stat Soc B 72(Part 1):3–25
39. Cadima J, Cerdeira JO, Minhoto M (2004) Computational aspects of algorithms for variable selection in the context of prin-cipal components. Comput Stat Data Anal 47:225–236
40. Dias LG, Sequeira C, Veloso ACA, Sousa MEBC, Peres AM (2014) Evaluation of healthy and sensory indexes of sweet-ened beverages using an electronic tongue. Anal Chim Acta 848:32–42
41. Söderström C, Rudnitskaya A, Legin A, Krantz-Rülcker C (2005) Differentiation of four Aspergillus species and one Zygo-saccharomyces with two electronic tongues based on different measurement techniques. J Biotechnol 119:300–308
42. Rudnitskaya A, Kirsanov D, Legin A, Beullens K, Lammertyn J, Nicolaï BM, Irudayaraj J (2006) Analysis of apples varieties— comparison of electronic tongue with different analytical tech-niques. Sens Actuators B 116:23–28
43. Kuhn M (Contributions from Wing J, Weston S, Williams A, Keefer C, Engelhardt A, Cooper T, Mayer Z, and the R Core Team) (2014) caret: classification and regression training. R pack-age version 6.0-24. http://CRAN.R-project.org/package=caret 44. Cerdeira JO, Silva PD, Cadima J, Minhoto M (2012) subselect:
selecting variable subsets. R package version 0.12-2. http:// CRAN.R-project.org/package=subselect
45. Mevik B-H, Wehrens R, Liland KH (2011) pls: partial least squares and principal component regression. R package version 2.3-0. http://CRAN.R-project.org/package=pls
46. Chung D, Chun H, Keles S (2012) spls: sparse partial least squares (SPLS) regression and classification. R package version 2.1-2. http://CRAN.R-project.org/package=spls
47. Venables WN, Ripley BD (2002) Modern applied statistics with S (Statistics and Computing), 4th edn. Springer, New York. ISBN 978-0-387-21706-2
48. Correia DPA, Magalhães JMCS, Machado AASC (2005) Array of potentiometric sensors for simultaneous analysis of urea and potassium. Talanta 67:773–782
49. Cimato A, Dello Monaco D, Distante C, Epifani M, Siciliano P, Taurino AM, Zuppa M, Sani G (2006) Analysis of single-culti-var extra virgin olive oils by means of electronic nose and HS-SPME/GC/MS methods. Sens Actuators B 114:674–680
50. Agiomyrgianaki A, Petrakis PV, Dais P (2012) Influence of har-vest year, cultivar and geographic origin on Greek extra virgin olive oils composition: a study by NMR spectroscopy and biom-etric analysis. Food Chem 135:2561–2568
51. Pouliarekou E, Badeka A, Tasioula-Margari M, Kontakos S, Longobardi F, Kontominas MG (2011) Characterization and classification of Western Greek olive oils according to cultivar and geographical origin based on volatile compounds. J Chroma-togr A 1218:7534–7542
52. Uncu AT, Frary A, Doganlar S (2015) Cultivar origin and admix-ture detection in Turkish olive oils by SNP-based CAPS assays. J Agric Food Chem 63:2284–2295
